ERNIE 4.5
ERNIE 4.5 is an enhanced version of ERNIE 4.0 with improved reasoning capabilities, better instruction following, and stronger performance across multiple domains including mathematics, coding, and creative writing. Features advanced knowledge integration, multilingual support, and optimized thinking capabilities. This 21B parameter model uses A3B quantization for efficient inference.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| AGIEval | 28.5% | self-reported llm-stats | link → |
| ARC-C | 40.6% | self-reported llm-stats | link → |
| ARC-E | 60.7% | self-reported llm-stats | link → |
| BBH | 30.4% | self-reported llm-stats | link → |
| C-Eval | 40.7% | self-reported llm-stats | link → |
| CLUEWSC | 48.6% | self-reported llm-stats | link → |
| CMMLU | 39.8% | self-reported llm-stats | link → |
| DROP | 28.6% | self-reported llm-stats | link → |
| GPQA | 74.0% | self-reported llm-stats | link → |
| GSM8k | 25.2% | self-reported llm-stats | link → |
| HellaSwag | 33.0% | self-reported llm-stats | link → |
| HumanEval+ | 25.0% | self-reported llm-stats | link → |
| MATH | 12.4% | self-reported llm-stats | link → |
| MBPP+ | 40.2% | self-reported llm-stats | link → |
| MMLU | 41.9% | self-reported llm-stats | link → |
| MMLU-Pro | 16.0% | self-reported llm-stats | link → |
| MMLU-Redux | 43.2% | self-reported llm-stats | link → |
| PIQA | 55.2% | self-reported llm-stats | link → |
| SimpleQA | 1.8% | self-reported llm-stats | link → |
| Winogrande | 51.3% | self-reported llm-stats | link → |